Introduction
This LoRA is trained on Akemi Homura (暁美ほむら), a character from Mahou Shoujo Madoka Magic (魔法少女まどか☆マギカ).
This LoRA was created at a friend's request, as they couldn't find a particularly good Homura model for Anima.
Usage
Recommended LoRA strength: 0.9-1.0
Trigger word: akemi homura
Magical girl outfit:
capelet, white shirt, frilled skirt, frills, pleated skirt, miniskirt, neck ribbon, purple ribbon, purple bow, black pantyhose, thigh bootsSchool outfit:
school uniform, juliet sleeves, puffy sleeves, white collar, red bowtie, red bow, plaid clothes, plaid skirt, pleated skirt, black skirt, miniskirt, black pantyhose, loafers, brown footwearSchool outfit (with glasses / Moemura):
(➕ TIP: Just append these tags to School outfit.)
glasses, semi-rimless eyewear, under-rim eyewear, twin braids, braid, hair bobbles, hair ornamentAkuma Homura:
(⚠️ NOTE: may not work very well. I'm going to train on more akuma homura images if having time)
akuma homura, mahou shoujo madoka magica, black dress, black wings, feathered wings, feathers, black gloves, elbow gloves, choker, bare shoulders, thighhighs, red ribbon, red bowTraining Details
Training tool:
sorryhyun/anima_lora v1.17.1Dataset: 21 manually captioned files
LoRA rank / alpha: 32 / 128
Learning rate:
2e-5Optimizer: AdamW
Learning rate schedule: Cosine decay with 5% warmup
Training duration: 600 steps
Precision: BF16
Resolution: Aspect-ratio buckets, approximately 1 megapixel
GPU: NVIDIA GeForce RTX 3090
简介
这个 LoRA 使用《魔法少女小圆》（魔法少女まどか☆マギカ）中的角色晓美焰（暁美ほむら）进行训练。
一位朋友一直找不到特别好用的 Anima 晓美焰模型，因此我应其请求制作了这个 LoRA。
使用方法
推荐 LoRA 强度： 0.9-1.0
触发词： akemi homura
魔法少女服装：
capelet, white shirt, frilled skirt, frills, pleated skirt, miniskirt, neck ribbon, purple ribbon, purple bow, black pantyhose, thigh boots校服：
school uniform, juliet sleeves, puffy sleeves, white collar, red bowtie, red bow, plaid clothes, plaid skirt, pleated skirt, black skirt, miniskirt, black pantyhose, loafers, brown footwear校服（戴眼镜）：
(➕ 提示：在校服提示词后追加以下标签即可。)
glasses, semi-rimless eyewear, under-rim eyewear, twin braids, braid, hair bobbles, hair ornament恶魔焰：
(⚠️ 注意：效果可能不太理想。如果有时间，我会用更多恶魔焰的图片继续训练。)
akuma homura, mahou shoujo madoka magica, black dress, black wings, feathered wings, feathers, black gloves, elbow gloves, choker, bare shoulders, thighhighs, red ribbon, red bow训练详情
训练工具：
sorryhyun/anima_lora v1.17.1数据集： 21 个手工标注的文件
LoRA rank / alpha： 32 / 128
学习率：
2e-5优化器： AdamW
学习率调度： 余弦衰减，预热阶段占 5%
训练时长： 600 步
精度： BF16
分辨率： 按宽高比分桶，约 100 万像素
显卡： NVIDIA GeForce RTX 3090
Description
Comments (4)
What is the point of this LoRA when Anima already knows Homura perfectly well? Now, if you had accurately replicated the original anime's style, many people would have been grateful. As it stands, this is a lazy, worthless hack job—it takes neither brains nor effort to scrape together a hodgepodge of fan art and throw it into a LoRA trainer.
This is a very strange and unreasonable accusation:
1. I think I already explained in the first paragraph of the description that Anima's depiction of Homura's features is not very stable (at least according to my friends' tests), and it often misses some clothing details, which is why I was commissioned to make this LoRA.
2. "Now, if you had accurately replicated the original anime's style, many people would have been grateful" - I think I already replicated the original anime's style in Demo Image 3? This LoRA is meant to replicate the style/character design of the characters in the original anime, so I have no idea what you're talking about.
3. "this is a lazy, worthless hack job - it takes neither brains nor effort to scrape together a hodgepodge of fan art and throw it into a LoRA trainer." - Wow, wow. Do you have any evidence? I think criticizing someone's work like this without any evidence is extremely irresponsible and leaves me speechless.
I'm not ready to fully disclose my LoRA creation process yet, but I don't think cherry-picking a few dozen of the better-quality images from tens of thousands of ones from the internet, then manually crafting captions one by one, training in parallel, and debugging hyperparameters is easy work.
You make it sound so easy; why don't you show everyone YOUR masterpiece?
4. Remember: all models are released for free. This is NOT a product, you have NOT paid me any money, you are NOT a "customer," and I have NO obligation to satisfy you. What an awful, arrogant, autistic mindset. Maybe you should go touch the grass.
Blocked.
@volda I am absolutely fed up with Civitai being flooded with low-quality LoRAs, so I don't hold back when I see them.
A proper LoRA requires at least 50 high-quality screenshots—not just random images scraped from booru's and Google.
I couldn't get the authentic Studio Shaft style out of your LoRA. The default prompt doesn't work here. Which one did you use? Why did you hide it, giving people yet another reason to complain or level accusations?
Anima knows Homura perfectly well; anyone capable of running it locally can verify that for themselves.
I’ve created LoRAs too—on a different account—specifically for Illustrious. I can send them to you. As someone who knows a thing or two about this, it really irritates me to see lazy LoRA work.
@Chichibonya Oh my God.
“I am absolutely fed up with Civitai being flooded with low-quality LoRAs, so I don't hold back when I see them” - So? I don't see how that has anything to do with my description. Wow! You have such a strong “sense of justice”; I imagine the people running Civitai would thank you too - but what does that have to do with my LoRAs? I don't think I'm making “low-quality LoRAs,” and I explained why, and you haven't offered any legit rebuttal.
“A proper LoRA requires at least 50 high-quality screenshots.” “Which one did you use? Why did you hide it, giving people yet another reason to complain or level accusations?” - What kind of bizarre standard is this now? Holy Jesus, I am completely and utterly speechless.
First of all, let me make this clear: Civitai does not require you to show the specific training information at all (it's all optional). The reason I list it is purely personal preference; I want to share training details and discuss training methods with others as much as possible. I never expected that information I made public out of goodwill would turn around and become a target for attacks against me. If Civitai is flooded with nitpickers like you, next time I simply won't list any information at all.
Second, technically speaking, almost all tutorials recommend using 10-30 images. More images won't make your LoRA quality better; image quality matters FAR more than quantity. Do you know how many images I collected for https://civitai.com/models/2937584/style-artist-cli? The answer is 361. However, after various rounds of cleaning, I ended up with only 61. Because I found that training them all mixed together actually degraded the results a lot, ultimately producing a bad mixed “average art style” state.
By the way, for this LoRA I collected more than 100 images together with a friend, then selected 21 higher-quality ones, and that was how we got the most satisfactory training results. But why should I mention these details here? Or, to put it the other way around, why? Because I showed the actual number of images used for training, I should be arrogantly accused by a total stranger who may never have even done grid training for a LoRA or compared results? Just from this one sentence alone, I can tell that you don't understand LoRA training deeply.
“I couldn't get the authentic Studio Shaft style out of your LoRA.” - Technical issue. I'm glad to see that you're the only one here raging impotently, while everyone else can use this LoRA normally.
“it really irritates me to see lazy LoRA work” - I won't repeat here the stuff about community etiquette (that is, without knowing any details at all, opening your mouth and arrogantly criticizing someone else's LoRA as “lazy” and “requiring no effort at all,” using only your own standards and assuming they are broad and universally applicable).
Fortunately, I do actually know a thing or two about LoRA training (I’d call myself beginner-to-advanced). I can't even imagine how violently enraged you'll be the next time you see a LoRA submitted by a “true beginner,” lmao.
And breaking news: there have ALWAYS been people in this world who are just starting to learn model training, and not all low-quality LoRAs come from malicious or lazy people (in fact, I'd say 90% of them don't). If Civitai were flooded with severely autistic people, we'd soon stop seeing any new models at all.
You could have asked me for more details in a normal tone and shown images to explain why you were dissatisfied with the LoRA. Maybe then we could have had a more constructive technical discussion, instead of you jumping straight to personal attacks.
Oh by the way, if my LoRA really gives you nightmares at 3 a.m. or makes you stub your toe while walking, just block me.